cross-agent MCP

Let your Claude Code, Codex, Grok Build and Copilot sessions talk to each other, live, without losing their existing context.

Copilot support targets existing VS Code Agent Host sessions, including sessions using custom models. It connects to VS Code's authenticated local Agent Host Protocol (AHP) endpoint; the model, tools and approvals remain managed by the existing host. Copilot does not use the CLI fallback described below and requires an exact session name, UUID or Copilot session URI.

cross-agent MCP is a relay MCP server for the coding agents you already have running — Claude Code, Codex, Grok Build and Copilot. Any one of them can hand a message to another through send_to_codex / send_to_claude / send_to_grok / send_to_copilot, keeping both sides' existing context. Claude, Codex and Grok work in VS Code or a plain terminal: the bridge finds their currently active session from the transcript on disk and resumes it. Their optional panel shims make the exchange render live in the real chat panel, as covered in section 3. Copilot requires VS Code and uses its existing Agent Host session directly.

Use send_to_copilot(session_id="copilot-test", message="...") for an existing Copilot conversation. list_agent_sessions(agent="copilot") includes its editor UUID, host URI, working directory and selected model. This integration has been verified with VS Code 1.141.0 / AHP 0.10.0 on macOS. Other AHP versions are refused until supported explicitly.

That's the difference from just running a second CLI by hand: neither side has to re-explain the task, and neither one loses the conversation it was already having. A few things this is useful for:

• Get a second opinion without leaving your conversation. Ask Codex to review or double-check Claude's plan, or the other way around, and keep working while it thinks.

• Hand off a long task and keep going. send_to_* is asynchronous — it queues the message and returns immediately. Whenever the peer's answer is ready, it arrives back as a new message in your own session — provided that session is open in a panel, or it is the Copilot caller's verified SDK session on its existing host. Otherwise the answer waits on the delivery record for bridge_status, and the receipt says so up front.

• Watch it happen, not just read a log. With the panel shims from section 3 installed, Claude, Codex and Grok exchanges render in VS Code's real chat panels. Copilot uses the host protocol directly and needs no shim.

• Long active turns keep running. Headless turns use an inactivity timeout, reset by CLI output or target transcript writes. Accepted panel turns are watched until they finish.

A real example

One user runs a single Claude session as a master that directs about a dozen role-based sub-sessions — a Claude or Codex session per role (implementation/tooling, image generation, client, server, copy, and a few more). The master only delegates work whose output would be too large for its own context; it keeps verification for itself, re-checking a peer's claims against the actual files or source lines rather than taking a self-report as done. Every delegation names an explicit session_id, since a dozen sessions sharing one working directory makes auto-selection unreliable, and handoffs between two Claude sessions use allow_same_agent: true.

In one run, the image-generation role (Codex) produced a request that the tooling role (Claude, same-agent) refused to act on — a field its contract required was missing. The master traced the fault to source, had the image role rewrite the request, verified the fix independently, and re-dispatched — three hops between two roles, with the master gating each one before it reached the human.

Contents

• 1. Requirements

• 2. Installation

• 3. Registration

• 4. Tools

• 5. Session resolution rules

• 6. Preventing infinite calls

• 7. Environment variables

• 8. Verification

• 9. Known limitations

• License

          VS Code or Terminal                 VS Code Agent Host
                   │                                  │
        ┌──────────┼──────────┐                       │
        │          │          │                       │
      Claude     Codex       Grok                   Copilot
     session A  thread B   session C                chat D
        │          │          │                       │
        └───────────── cross-agent MCP ────────────────┘
                              │
                      session registry
               (~/.cross-agent/registry.json)

Any of these can reach any other through the same hub — across products (Claude ↔ Codex ↔ Grok ↔ Copilot), or between sessions within the same product (see the same-agent rows in the table below, and section 6 for how that's gated).

¹ Same-agent rows need an explicit target — allow_same_agent=true or a session_id for Claude and Grok, a session_id for Codex. Copilot always needs session_id or a sticky pin; using that pin for Copilot → Copilot also needs allow_same_agent=true — see section 6.

With the shim attached, the exchange renders directly in the real VS Code panel (see section 3).

The exchange is asynchronous. send_to_* queues the message and returns immediately — it does not carry the peer's reply. A background worker runs the peer's turn, and once a reply exists, it is delivered as a new message through the sender's live return route (see Return traffic needs a panel, including Copilot's verified SDK route). Because nothing blocks, neither session is locked while the peer's turn runs, and a turn that takes several minutes won't be lost to a timeout.

send_to_codex ──▶ [outbox queue] ──▶ Codex turn (minutes)
     │                                   │
returns immediately                answer generated
  (delivery_id)                         │
                                        ▼
                  delivered to the Claude session as a "BRIDGE REPLY" message

Reply address

Since the reply needs somewhere to land, the sender must know its own session precisely. For a per-conversation panel process, the MCP server is a descendant of its own shim, so the shim in its ancestor chain identifies that conversation. Shared runtimes need the caller bindings described below.

A pin must not substitute for this. A pin records "where to send," not "who I am." When it once worked that way, a stale pin got used as the reply address and replies landed in the wrong session.

Grok Build hands it over directly: it exports GROK_SESSION_ID to every MCP server it starts, and starts one per session — even when several sessions share one grok agent stdio process (checked by opening two in one process: two servers, each with its own id). So a Grok caller's return address is exact without any inference. The variable is inherited by everything Grok runs, which is why it counts only for a Grok caller: a claude started from one of Grok's shell commands is not that session.

Codex needs one more level of precision. All the Codex threads in one window share a single app-server and a single MCP server, so the process tree can only tell you "this window," not which thread within it (back when the most-recently-active thread was picked instead, four replies landed in threads that had never asked anything). Instead, Codex attaches x-codex-turn-metadata (thread id, turn id) to every MCP call, so the bridge uses the id the calling thread declares about itself as the reply address. When this value is present it takes priority over inference. Claude Code runs a separate process per conversation, so the process tree alone is enough there.

Copilot's SDK starts a separate MCP child per SDK session and supplies COPILOT_AGENT_SESSION_ID. The bridge maps that SDK UUID to the editor's Host UUID through registered default-chat metadata. It never uses a target pin or a shared runtime PID as the caller address. An invalid SDK binding or a conflict stops a send; see Copilot registration.

The address is also written into the envelope — like the From line of an email.

=== CROSS-AGENT BRIDGE MESSAGE ===
from: Claude Code (peer AI agent, not the human user)
reply-to: claude session 058a16bc-3a77-4604-a328-9409c391f918
conversation: conv_7bc3806dc3a8 | hop 1/4

Since the bridge delivers the reply on its own, this line doesn't establish the reply. It earns its keep when the automatic path fails, and when the peer sends back a new request — it can target that exact session without re-inferring what's active on this side.

Reply delivery runs from the sender session's directory. Claude transcripts are stored under their own project directory, so a delivery aimed at the directory the request was headed toward produces No conversation found even when the session itself is fine. (A reply is never delivered by resuming the session — see below — but it still carries that session's directory, because a notice and a reply are built the same way and the directory is what makes the address complete.)

Return traffic needs a panel

A request is addressed: somebody chose a session and sent to it, and resuming that session over the CLI is the delivery they asked for. Return traffic is not addressed. A reply, and the DELIVERY FAILED notice that stands in for one, goes back to whichever session started the exchange — and that session was live moments ago, because it was the one that sent.

Return traffic is delivered through that session's live editor transport. For Claude, Codex and Grok, this requires a panel. Copilot can also receive it on its caller's own verified SDK session while the chat is hidden; the exact binding and host ownership are checked again before dispatch. Return delivery never resumes a CLI or starts another SDK runtime.

When the sender has neither an available panel nor that verified Copilot return route:

• the send_to_* receipt says so at once — return_panel_available_now: false, plus a warning and a note saying where to read the answer instead. It reports the probe, not a promise: the route is resolved again when the answer exists, so a panel opened in the meantime is used, and one closed in the meantime is not replaced by a resume;

• bridge_status(delivery_id=...) reports is_return_status_only: true, keeps a 2,000-character reply_preview on the record, and reads the full answer fresh from the peer's transcript as peer_transcript.answer. If that transcript is later gone, the preview is what remains. Nothing is re-sent;

• a reply whose panel closes after the route was resolved is retried — a reopened tab is found again — and fails as undelivered rather than resuming. The request names that delivery as return_delivery_id, and bridge_status on the request reports its outcome under return_delivery: a request finishes before its answer is delivered, so on its own it cannot say whether the answer landed.

Sending to a dormant Claude, Codex or Grok session still resumes over the CLI, and the receipt warns when it has been idle long enough to look retired. Copilot requires a reachable existing host even when the target chat is hidden.

Related MCP server: claude-code-relay

1. Requirements

• macOS / Linux, Python 3.10+

• For each CLI product you relay to: claude (Claude Code 2.x), codex (0.146+), or grok (Grok Build, developed against 1.0.44)

• Every CLI you use logged in (grok login, or XAI_API_KEY — see What a spawned agent CLI inherits)

• For Copilot: VS Code with an existing Copilot Agent Host default chat, including a chat using a custom model. Verified on macOS with VS Code 1.141.0 / AHP 0.10.0; no standalone Copilot CLI is required. See registration.

• For Grok in a VS Code chat panel (optional; everything else works from a terminal): the Grok Build GUI extension (SahilRakhaiya.grok-build-gui, developed against 1.0.4). It is the only Grok extension this has been verified with — see IDE panel integration

2. Installation

cd ~/project/cross-agent_mcp
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
chmod +x run-server.sh

3. Registration

Register it in each agent at the user (global) level.

Claude Code

claude mcp add cross-agent -s user \
  -- ~/project/cross-agent_mcp/run-server.sh

This is written into the top-level mcpServers of ~/.claude.json, so it's available in every project. To attach it to just one project use -s local; to share it via the repository, use a project-root .mcp.json.

{
  "mcpServers": {
    "cross-agent": {
      "command": "~/project/cross-agent_mcp/run-server.sh"
    }
  }
}

Codex

codex mcp add cross-agent \
  -- ~/project/cross-agent_mcp/run-server.sh

This adds the following to ~/.codex/config.toml (Codex only supports global config).

[mcp_servers.cross-agent]
command = "~/project/cross-agent_mcp/run-server.sh"
default_tools_approval_mode = "approve"   # so the UI doesn't show an approval prompt every time (added manually)

default_tools_approval_mode has no corresponding flag on codex mcp add, so it's added directly to config.toml. Valid values are auto / prompt / writes / approve; use approve to stop the approval prompt from popping up every time (auto kept asking). This does not fix the cancellation problem with headless codex exec (see section 9). Clicking "Always allow" once on the UI prompt has the same effect.

Copilot (VS Code Agent Host)

Register this stdio server in VS Code's user-level <VS Code user data>/User/mcp.json (macOS Stable: ~/Library/Application Support/Code/User/mcp.json; Linux Stable: ~/.config/Code/User/mcp.json):

{
  "servers": {
    "cross-agent": {
      "type": "stdio",
      "command": "/absolute/path/to/cross-agent_mcp/run-server.sh"
    }
  }
}

For Copilot turns sent while no editor client is attached to that conversation, also merge the same definition into the Agent Host's persistent mcpServers map in <VS Code user data>/User/globalStorage/agent-host-config.json:

{
  "mcpServers": {
    "cross-agent": {
      "type": "stdio",
      "command": "/absolute/path/to/cross-agent_mcp/run-server.sh"
    }
  }
}

Preserve the file's other settings and other server definitions. On macOS Stable the file is ~/Library/Application Support/Code/User/globalStorage/agent-host-config.json; on Linux Stable it is ~/.config/Code/User/globalStorage/agent-host-config.json.

After adding this host definition to a running installation, fully quit and reopen VS Code once. The host loads this file at startup; Reload Window can keep the same host process alive and does not reread it. Its supported root/configChanged API can apply and persist the same declaration without restarting, but the MCP server does not silently change host settings. User mcp.json alone is an editor client contribution: a reload or detached chat can remove it from an existing SDK session. The host-level definition keeps cross-agent available there. An active client's definition with the same server name takes precedence.

No shim or extra extension is installed. The bridge discovers live editor endpoints under <VS Code user data>/agent-host/local-endpoint/entries/, authenticates using the existing owner-only metadata and subscribes to the target session's exact default chat. Authentication tokens never enter bridge records or logs. Both the outgoing model selection and its options are preserved from that chat. Accepted turns are correlated by their own turn ID and client sequence; an uncertain hand-over is reported without resending it.

Copilot must already have an existing session on a reachable host. new_session=true, a missing target and a closed/unreachable host are refused; no CLI runtime is started. This does not cover legacy extension-host Copilot Chat or standalone Copilot CLI sessions. Returns to Copilot go through the same existing host, never a second runtime. Read-only recovery links the editor's UUID to the separate Copilot SDK session UUID in its metadata and reads completed SDK events matching the delivery token.

Copilot's SDK gives its MCP process COPILOT_AGENT_SESSION_ID, the exact SDK session UUID. The bridge maps it through registered defaultChatProviderData.sdkSessionId metadata to the editor's Host UUID, including for bridge_status. It does not choose the newest chat, use a target pin as the caller, or derive a session from the shared runtime PID. The inherited variable is ignored when another known agent (such as Codex or Claude) is the caller.

There must be one unarchived registered default chat for that SDK UUID. Missing metadata, an archived session, a side-chat/subagent UUID without a default-chat mapping, or duplicate mappings stop a send instead of inventing a return address. If both MCP caller metadata and the SDK binding are present, they must name the same Host UUID. Older hosts supplying neither an exact caller URI nor the SDK variable retain status-only return traffic.

Copilot is detected from its runtime process or its own SDK variable, so this registration does not force CROSS_AGENT_SELF=copilot on servers that another Agent Host runtime may use. The SDK caller binding is not forwarded to peer CLIs started by the bridge.

Tool approvals are separate from session delivery. With Manual permissions, Copilot may ask before using cross-agent tools. For the tested Agent Host version, settings.json and mcp.json do not provide a verified per-server persistent auto-approval declaration. chat.tools.eligibleForAutoApproval only controls eligibility. The supported setting chat.tools.global.autoApprove applies to all tools; it is not a cross-agent-only permission.

If the installer explicitly chooses global automatic approval, add the following to VS Code's User settings.json, then run Developer: Reload Window once:

"chat.tools.global.autoApprove": true

This includes file edits, terminal commands and other tools, as well as cross-agent. It also affects existing Manual-permissions Copilot Agent Host sessions. A workspace-only setting does not provide the same Agent Host behavior. Leave it unset to retain the existing approvals. The MCP server itself never changes these permission settings or replays a skipped tool call. See VS Code approval settings.

A stored catalog entry does not establish a return address. A reply or failure notice requires an attached editor client, or the caller's own verified SDK session on its existing host. The latter allows a Copilot caller to receive an answer while its chat is hidden: its SDK UUID must resolve to exactly that default chat, and the authenticated endpoint owner must still be an ancestor of the MCP process. Both checks are repeated immediately before return dispatch. The bridge never starts another SDK runtime or resumes a CLI to deliver the answer. Its read-only observer subscription does not count as an editor client. Without either proof, the answer remains available through bridge_status.

Default paths cover VS Code Stable and Insiders on macOS/Linux. For a custom --user-data-dir, set these path-list environment variables (separate multiple paths with :):

The host DB and session directories must correspond in order when arbitrary override paths are used. Duplicate UUIDs in different configured stores are refused rather than selecting the newest copy. These paths are always read-only.

Grok Build

Grok imports the MCP servers already registered for Claude Code (~/.claude.json), so once the Claude registration above exists, a Grok session started afterwards already has cross-agent. To register it for Grok on its own:

grok mcp add cross-agent \
  -- ~/project/cross-agent_mcp/run-server.sh

This adds the following to ~/.grok/config.toml; a server named the same in both places is one server, and config.toml wins.

[mcp_servers.cross-agent]
command = "~/project/cross-agent_mcp/run-server.sh"

Grok reaches an MCP tool through its own search_tool / use_tool, so the model finds send_to_grok and the others by searching, and asks for approval on the first use in an interactive session.

Unattended use

With the commands above, a session the bridge starts fresh stays at the safe defaults: read-only for a new Codex session, and Claude's own default permission mode. That means the first edit or command either agent wants to make stops and asks — including inside a turn the bridge itself started, where there's no human sitting there to answer it.

If you don't want to deal with that prompt, add these two variables to the same registration:

claude mcp add cross-agent -s user \
  -e CROSS_AGENT_CODEX_SANDBOX=danger-full-access \
  -e CROSS_AGENT_CLAUDE_PERMISSION_MODE=bypassPermissions \
  -- ~/project/cross-agent_mcp/run-server.sh

codex mcp add cross-agent \
  --env CROSS_AGENT_CODEX_SANDBOX=danger-full-access \
  --env CROSS_AGENT_CLAUDE_PERMISSION_MODE=bypassPermissions \
  -- ~/project/cross-agent_mcp/run-server.sh

Grok has the same switch, and it is not part of the two above: CROSS_AGENT_GROK_PERMISSION_MODE is passed to grok as --permission-mode (acceptEdits, bypassPermissions, …).

grok mcp add cross-agent \
  -e CROSS_AGENT_GROK_PERMISSION_MODE=acceptEdits \
  -- ~/project/cross-agent_mcp/run-server.sh

Unset, a headless Grok turn cancels the first tool call that needs an approval — nobody is there to give it — and the delivery fails with an error that says so, instead of handing back half a sentence ("let me run that") as the answer. Read-only tool calls run without asking in every mode, so questions, reviews and reads need nothing set. acceptEdits lets it write files; bypassPermissions lets it run anything, and carries the same warning as the two variables above. It applies to new and resumed sessions alike.

They affect only the sandbox and permission mode of an agent the bridge starts. The bridge's own tool calls have a separate approval setting (default_tools_approval_mode above for Codex, "Turning off the approval prompt" below for Claude Code), and panel integration needs neither.

Turning off the approval prompt (Claude Code)

Claude Code asks for approval on every MCP tool call. Add a server-level rule to ~/.claude/settings.json (a single server name, not a per-tool list or a * wildcard).

{
  "permissions": {
    "allow": ["mcp__cross-agent"]
  }
}

Cross-session inbound approval (native SendMessage, not this bridge)

Separate from the mcp__cross-agent__* tools above, Claude Code also ships its own built-in cross-session messaging (SendMessage / ListAgents), unrelated to this repo's code. When one Claude Code session messages another this way, the recipient applies a permission-mode check:

• Unset (default): the message auto-delivers only when the sender's permission-mode class matches the recipient's (bypassPermissions↔bypassPermissions or prompting↔prompting). A mismatch — e.g. the recipient runs bypassPermissions (see "Unattended use" above) but the sender doesn't — holds the message for the recipient's human to approve before Claude ever sees it.

• To skip that hold, set on the recipient session, in its .claude/settings.json:

  {
    "crossSessionInbound": "accept"
  }

  Other values: "hold" (always require approval, even on a mode match) and "refuse" (opt the session out of inbound cross-session messages entirely).

IDE panel integration (bidirectional)

This section covers the optional Claude, Codex and Grok panel shims. Those products also work from a plain terminal. Copilot uses the Agent Host transport described in its registration section and needs none of these shims.

The CLI resume path (codex exec resume / claude -p --resume) appends a turn to the session history, so context is preserved, but it doesn't show up in the VS Code panel. The panel's session lives only inside the child process the extension spawned and connected to directly over stdio, and there's no way in from outside.

Inserting the shims into the middle of that pipe solves it.

VS Code extension ──stdio──▶ codex-shim.sh  ──stdio──▶ real codex app-server
VS Code extension ──stdio──▶ claude-shim.sh ──stdio──▶ real claude (stream-json)
VS Code extension ──stdio──▶ grok-shim.sh   ──stdio──▶ real grok agent stdio (ACP)
                                  ▲
                                  │ unix socket
                          cross-agent MCP  ──▶ inject message ──▶ rendered in panel

Add these to VS Code user settings and reload the window.

"chatgpt.cliExecutable": "~/project/cross-agent_mcp/codex-shim.sh",
"claudeCode.claudeProcessWrapper": "~/project/cross-agent_mcp/claude-shim.sh",
"grok.cliPath": "~/project/cross-agent_mcp/grok-shim.sh"

Shared rules:

• Passes every byte straight through, and intercepts only panel-session calls (--version, login, app-server daemon, claude -p, grok update, grok agent serve, etc. exec straight to the real binary)

• Auto-discovers the real binary inside the extension directory (Grok: its own install, ~/.grok/bin/grok) — can be overridden with CROSS_AGENT_REAL_CODEX / CROSS_AGENT_REAL_CLAUDE / CROSS_AGENT_REAL_GROK

• On any failure, it execs the real binary as-is (fail-open)

• The shim records its own pid ancestor list in ~/.cross-agent/panels/<agent>-<pid>.json. The bridge picks the shim that shares an ancestor with itself, so even with multiple windows open it targets exactly "this IDE instance"

• The Claude shim waits for the current turn to finish before injecting, if the user is mid-conversation

• The Grok shim does the same per session, and keeps count of the human's prompts still unanswered: Grok queues a prompt typed ahead, so the session is busy until the last one is answered, not the first

• The Grok shim answers for every session its process hosts. The extension starts one grok agent stdio per conversation tab, but a process can carry several sessions, and each is reported and addressed by its own id. A session the extension opened and nobody has prompted yet counts as empty — a message written into it is the new conversation — while one that has history is never used for a new conversation, because the conversation list belongs to the extension and a session made behind its back would be one the panel never heard of

• What the Grok panel shows. The extension draws the messages it sends itself and discards the agent's echo of a prompt unless it is replaying a history, so a user bubble cannot be made from outside. The shim instead writes the relayed message as a quoted block at the head of the reply, and Grok's answer follows it. It also holds back the Thinking blocks of a bridged turn — the extension opens one for each and closes it only when its own prompt is answered. The panel's own busy state and Stop button follow only prompts it sent, so they do not light up for a bridged turn; a human who types meanwhile is queued by Grok, and the shim reports the session busy until that turn ends too

• Only one Grok extension has been verified: Grok Build GUI (SahilRakhaiya.grok-build-gui, 1.0.4). Everything the shim does about the panel — what it hides, what it shows, which messages the extension throws away — was read from that extension's code and checked against a stand-in for it, so it is that extension's behaviour and nobody else's. A different Grok front end that launches grok agent stdio from a path you can configure could be pointed at grok-shim.sh the same way, but that is untested and its panel may draw the same messages differently. One that talks to Grok another way (a socket to grok agent serve, say) is not intercepted, and its sessions are reached over the CLI. The CLI path, session discovery and the send_to_grok tool do not depend on any extension

• The Codex shim excludes sub-agent threads from targeting. Threads created by a multi-agent run reject direct input at the app-server level (direct app-server input is not allowed for multi-agent v2 sub-agents), and are identified via parentThreadId · agentNickname · agentRole · canAcceptDirectInput. A thread id arriving only in a notification is never enough on its own to create a new target — only thread/start · thread/resume · turn/start · turn/steer sent directly by the extension are trusted. If it's still rejected, the shim drops that thread, opens a new conversation, and retries once

A new conversation is the last resort

If a new conversation opens mid-task, all context up to that point is gone. The peer suddenly appears to remember nothing, so if there's any recoverable conversation at all, a new one is never created.

1. The session named by session_id        ← id or conversation name
2. The session pinned via pin_agent_session
3. The conversation open in this window's panel   ← visible directly in the panel
4. The active session on disk (CLI resume)  ← not visible in the panel, but context is intact
5. Only when none of these exist → a new conversation

The order of 3 and 4 matters. It used to be that "if the panel has no conversation, open a new one" fired before step 4, so a perfectly fine session on disk would still get a new conversation created over it.

You can specify by name — but only an exact match. People refer to conversations by name, not uuid, so you can put the conversation name directly into session_id.

• The name a human assigned is the source of truth. It's the name shown at the top of the panel, recorded in the transcript as {"type":"custom-title","customTitle":"…"}. Renaming appends another one, so the last value is used. For Codex threads, ~/.codex/session_index.jsonl supplies the name.

• Grok Build keeps two kinds of name. The tab name a human types in the Grok Build GUI extension lives in the extension's own state, not in Grok's files, so the bridge reads it, read-only, from the editor's state.vscdb (macOS and Linux locations for Code by default; CROSS_AGENT_GROK_NAMES_DB points somewhere else). A title set with /rename in Grok itself is in the session's summary.json. If the database cannot be read, names simply stop resolving and the session is reached by its id.

• A conversation with no assigned name falls back to a title generated from the first message. That's a description, not a name, so it can't be found by a word inside it.

• No partial matching. It used to allow it, and koppa_studio once matched a path quoted in a months-old session's first message, headlessly reviving a session nobody was watching — while the session actually named koppa_studio went unfound.

• If no name matches, it reports similar titles and fails rather than creating a new conversation. session_id and new_session can't be used together (they express opposite intents), and neither can session_id and pin_agent_session.

send_to_claude(message=..., session_id="studio_v4_orginial")
pin_agent_session(agent="claude", session_id="studio_v4_orginial")

Whether by name or id, if what was specified doesn't exist, it errors instead of creating a new one. Failing is better than silently starting a different conversation.

When a new conversation is opened, the response's warning field carries that fact and the reason.

Name your sessions

Addressing by name is only as reliable as the names are. To target the right session every time:

• Give every session that another session will send to a name of its own. Don't rely on the title generated from its first message: that is a description, not an address, and it isn't unique — the same opening prompt produces the same title, so one generated title can be shared by hundreds of sessions. A session the bridge creates has no name either, so name it as soon as it exists.

• Never use the same name for two sessions or threads. The bridge cannot tell them apart, and will not guess: a send or a pin that names a title two conversations answer to is refused, and the error lists the candidates - each with when it was last active and whether its name was assigned or generated from its first message - so you can address the one you meant by its id.

• When a session is retired or replaced — for example because its context is full — rename the old one to something different, and give the replacement the name. For instance, rename billing-api to billing-api-old-1 and name its successor billing-api. A retired session stays on disk and a lookup by name still finds it, so for as long as it keeps the live session's name, that name means two conversations.

• Lookup by name considers at most the 500 most recently active sessions, and for Codex it can stop sooner still, at the rollout-file bound CROSS_AGENT_CODEX_SCAN_LIMIT. So an old session may not be found by name at all — and a name matched only once by a search that stopped at either bound cannot be shown to be unique either, so it is refused rather than resolved on a maybe. The error says how many sessions were actually reached. If a name can't be kept unique and current, address the session by its id, or pin it (pin_agent_session).

When the panel has no conversation open

Even when the panel is only showing a conversation list, there's a live process behind it. Falling back to the CLI here means the requester gets an answer, but the panel stays empty, making it look like the bridge did nothing. So the shim opens a new conversation in the panel and puts the message there instead.

• Codex: creates a thread with thread/start. The app-server broadcasts a thread/started notification, so the extension picks up the thread and renders it

• Claude: just writes a user message to the panel process that's running without a session. The CLI starts a new conversation and the session id is captured from system/init

• Grok: writes the prompt into a session the extension opened and nobody has prompted yet. A shim cannot open a session of its own (see above), so when every session it hosts already has history, a request for a new conversation goes to the CLI instead: grok --single=<message> --session-id <new uuid>

Copilot never creates a conversation through the bridge. Create its chat in VS Code first, then send to its exact name, UUID or session URI.

When the receipt's will_create_session is true, this is the conversation that will be opened. Codex also sets a title of the form "sender: start of the message" via thread/name/set; otherwise it would just sit in the list as "New chat" with no way to tell which conversation it is.

The new conversation shows up in the list with an unread marker, but the panel doesn't automatically open it. The app-server protocol has no notification that moves the client to a specific conversation, and the extension's vscode:// deep link (the /local/<thread-id> route) can't target a specific window — it moves the Codex panel of every open VS Code instance at once. So it wasn't adopted.

Which conversation tab it goes to

The selection order below applies to Claude, Codex and Grok shims. Copilot requires an explicit target or a sticky pin and does not select a chat from focus or recent activity.

The extension spawns a separate process per conversation tab, so a single window ends up with multiple shims running. Nothing records which tab has focus, so it's picked using the following order of evidence.

1. The session explicitly given via send_to_*(session_id=...)
2. The session pinned via pin_agent_session
3. The tab a human typed into most recently (observed directly by the shim on the extension→agent path)
4. (If nobody has typed since the shim started — e.g. right after a window reload)
   the tab whose transcript was updated most recently
5. The most recently opened tab

Observed human input always takes priority over transcript timing. Turns the bridge itself injects also touch the transcript, so without this rule the bridge would keep re-picking the tab it last wrote to. Injected turns aren't counted as observed input, so this contamination never arises in the first place.

bridge_status's ide_panels shows the list of open tabs and the selection result as-is. If it's not the tab you want, pin one with pin_agent_session.

A send with no session_id is addressed by the human, not by you.

Rules 3 to 5 pick from what the person at the keyboard is doing. That is what you want for "ask Codex about this" while they watch. It is not an address: it moves when they switch tabs, so two sends in a row can land in different conversations, and a reply that arrives minutes later belongs to whichever tab was in front at the time. A status update in an ongoing exchange has been delivered into an unrelated thread this way, which then began acting on it.

Pass session_id explicitly for anything automated, delayed, or part of an exchange that continues over several messages — every time, not only the first. A long conversation with one peer starts to feel like an addressed channel and is not one.

The receipt tells you which happened: target_selected_by is caller when you named the session, pin when a pin set earlier did, and panel-focus or discovery when nobody did — caller_supplied_session_id is true only for the first, because a pin is standing configuration rather than a choice this call made. An unaddressed relay also returns a warning. Check target_session_id is the conversation you meant before reporting a send as done; a misdelivered message cannot be recalled.

To have such a send refused rather than warned about, set CROSS_AGENT_REQUIRE_EXPLICIT_TARGET=1 in both registrations — each client runs its own server, so the setting on one side guards only the sends made from that side. Every send then has to say where it goes itself: a session_id (an id, or a name that matches exactly) or new_session=true. Anything else is refused before a target is looked up, so nothing is sent and no hop is spent, and the error says what to pass instead. A pin does not count: it was set at some earlier point and says nothing about whether this call meant to go there, so a caller relying on one has to pass session_id as well. A blank or whitespace-only session_id counts as none. The mode is off by default, because "ask Codex about this" while you watch is exactly what panel focus is good for.

Conversations in another VS Code window

The shim socket is an ordinary unix socket and isn't tied to a window. What process-ancestor detection determines is "which window," not "can it be reached." So the rule splits into two.

bridge_status's ide_panels.<agent>.other_window_sessions shows conversations open in other windows. They're not candidates for auto-selection, but they're reachable if named explicitly via session_id.

Without this distinction, a session in another window would fall back to a headless CLI resume, which the CLI rejects with thread-store conflict: already has an active writer if that window's live panel is still holding onto the conversation. In other words, reaching outside the window for an explicitly named session isn't a convenience — it's the only path that actually succeeds.

CROSS_AGENT_UI_HOOK selects the behavior — auto (default: use it if present, fall back to CLI otherwise), off (always CLI), require (fail instead of silently falling back if the panel can't be found).

For Copilot, auto and require use its existing Agent Host endpoint; off refuses delivery because there is no CLI fallback.

Never starting a headless CLI

require only refuses a send that names no session (or a new conversation no panel can open). A session named by session_id (or by a pin) that no panel holds is resolved from the session store and resumed over the CLI — a second writer beside the editor's own panel, which stops taking instructions into that session while the turn runs. To rule that out entirely, set CROSS_AGENT_DISABLE_HEADLESS=1.

With it set, a send is delivered into a session an editor panel holds, or it is refused with nothing sent: no CLI is started, nothing is queued and no hop is spent, and the error says which session no panel holds. It is enforced twice — send_message refuses before anything is queued, and _run_cli, the one place a peer CLI is ever started, refuses again, so no route (a pin, discovery, a stale resolution) can step around it. bridge_status reports settings.disable_headless.

Each client runs its own server, so set it in every registration — the Claude Code one and the Codex one — for it to guard sends made from both sides. Anything not open in a panel is unreachable while it is on: open the conversation in its panel first, or leave the setting off for that run. CROSS_AGENT_UI_HOOK=off together with it refuses every send.

Status check

./run-server.sh --check

Prints who's running it, CLI paths, resolved sessions and the settings in effect, then exits. Run with no arguments, it comes up as an MCP stdio server and waits on stdin (this is normal — exit with Ctrl-C).

Applying code changes without restarting the editor

Every editor session starts its own server, and Python reads the source once, at start. So run-server.sh started with no arguments — which is how every registration above starts it — runs the server behind a small front: the editor talks to the front, and the front relays to the server (the worker), which is the ordinary server started as its child.

./run-server.sh reload

asks every running front to swap its worker for one started from the current source, then prints what each did:

reload 1792633199847-69436: asked 3 running front(s)
  ok      front 68063: server 68071 -> 69694 (generation 2)  cwd=/Users/you/project
  FAILED  front 70112: the new server (pid 70990) failed: ...; the old one keeps serving
  -       front 70544: the editor has not connected yet; nothing to replace

• Nothing in flight is cut off. New calls go to the new worker only after it has answered the editor's own handshake (initialize, or server/discover on MCP 2026-07-28). The old worker answers the calls it already took, then is told to drain: it stays until every delivery it carries — a peer turn it is watching, an answer on its way back — is done, and exits.

• A broken tree is refused before any front sees it. The command imports the source in a fresh process first; one that does not import exits with reload refused and nothing is asked of the fronts. A worker that still fails to start, or dies while taking over the editor's subscription, is reported per front as FAILED, and that front goes on with its old worker. The command exits non-zero when any front failed or did not answer in time.

• A call the editor cancels is not waited for. MCP servers never answer a cancelled request, so the front stops counting it as owed; an interrupted call does not keep a replaced worker alive.

• Closing the editor: the current worker exits as a server always has. A worker a reload already released goes on until its last delivery is done (at most 24 hours), still holding the busy locks of what it carries under its own pid; stopping the server with a signal stops it too.

• The editor keeps one connection. On MCP 2026-07-28 (Claude Code) the editor holds one subscriptions/listen request open for the whole session; the front moves it to the new worker under the editor's own request id, so it stays a single stream. After a reload the editor is told the tool list may have changed: Claude Code fetches it again, so a new tool or argument shows up at once. Codex does not refetch, so a change to a tool's signature reaches a Codex thread only when it starts a new server; changes to behaviour reach it at once.

• A worker that dies is replaced. Calls it had not answered are answered with an error, and the front starts another worker. One that keeps dying (more than 3 times in 60s) takes the front down with it, as a crashed server always did.

• What still needs an editor restart: changes to the front itself (front.py), to run-server.sh, to the server instructions (the editor reads them once, from the first handshake), and to environment variables of a registration.

bridge_status reports server_process: the worker's pid, its front_pid and its worker_generation, so a session can see which code is answering it. Set CROSS_AGENT_RELOAD=0 in a registration to run the server directly, without a front.

4. Tools

Common send_to_* parameters:

send_to_claude, send_to_grok and send_to_copilot additionally take allow_same_agent (default false) — a session messaging another session of its own agent is refused unless this is set, or an explicit session_id is given. send_to_codex has no such flag; reaching another Codex thread requires an explicit session_id (see section 6). For Copilot, the flag does not enable automatic target discovery: a sticky pin or explicit target is still required.

The return value of send_to_* is a receipt, not an answer.

5. Session resolution rules

Copilot only uses an explicit session_id or a sticky pin created with pin_agent_session. Its existing default chat must resolve uniquely and have a reachable authenticated host. It does not auto-select by activity, create a session, or fall back to a CLI. The scanning and creation rules below apply to Claude, Codex and Grok.

On a send_to_* call, the target session is determined in the following order. Under CROSS_AGENT_REQUIRE_EXPLICIT_TARGET only rule 1 applies, alongside new_session=true: a call with neither is refused before any of this runs, even when a pin exists (see Which conversation tab it goes to).

1. If a session_id argument is given → that session
     - Searched in the order: this window's panel → another window's panel → transcript on disk
2. If the registry has a pin → that session
     - A pin set via pin_agent_session never expires
     - A pin the bridge created automatically is valid only within the active window (default 240 min)
3. Scan the session store → the best-fit transcript matching the conditions
     - Claude: ~/.claude/projects/<slug(cwd)>/*.jsonl
               has a user message and isn't sidechain-only
     - Codex:  ~/.codex/sessions/**/rollout-*.jsonl
               session_meta.thread_source == 'user' (sub-agent threads excluded)
               if multiple rollouts share a session_id, the newest one
     - Grok:   ~/.grok/sessions/<percent-encoded cwd>/<session-id>/summary.json (+ updates.jsonl)
               session_kind != 'subagent' (a subagent's conversation is filed in the same store)
               the folder is named for the directory, so a scope is decided without opening sessions
     - Only counted as "active" if its last record falls within the active window (default 240 min)
     - Sort priority: ① directory match (exact > subdirectory > parent) ② most recently recorded
       Parent directories are only candidates under scope='tree' — since the home directory
       is a parent of every project, a session opened at ~ must not be able to hijack an
       arbitrary project
4. If no session satisfies the above → create a new session
     - Claude: claude -p --session-id <new uuid> ...
     - Codex:  codex exec --json -C <cwd> ... (id recovered from thread.started)
     - Grok:   grok --single=<message> --session-id <new uuid> ...
     - The created session gets pinned in the registry, becoming the resume target from the next call on

In short: if there's an active session, it continues the context; if not, it creates one and reuses it from then on.

6. Preventing infinite calls

The bridge applies these guards.

1. Hop budget — at most MAX_HOPS (default 4) per conversation. conversation_id propagates to child processes as an environment variable, so an A→B→A→B chain automatically shares the same budget. Rejected once exceeded.

2. Busy lock — prevents two turns from overlapping in the same session. Agent CLIs allow only one writer per transcript, so this isn't courtesy, it's mandatory. It's acquired atomically in ~/.cross-agent/locks/ with O_EXCL, so there's no race between "check" and "acquire." It's released only while its own token is still present, and locks held by dead processes are reclaimed automatically. Since the move to async, this lock is a reason to wait in line, not a reason to reject — the worker waits for the lock to free up and then delivers. The sender's own session is no longer locked at all, since nobody's waiting on it, so a relay coming back around can't create a deadlock.

3. Self-call guard — by default, an agent refuses to relay into another session of its own kind: Claude cannot reach another Claude session, Codex cannot reach another Codex thread, Grok cannot reach another Grok session, and Copilot cannot reach another Copilot chat. The caller is identified from process ancestry and host-supplied caller bindings. This exists because the peer tools (send_to_codex / send_to_claude / send_to_grok / send_to_copilot) are meant to cross from one product to another — routing Claude into Claude by mistake should fail loudly instead of silently starting a same-agent relay. The refusal names the tools that lead elsewhere.

   • Claude → Claude is allowed either by passing allow_same_agent=true on send_to_claude (auto-discovers another active Claude session, excluding the caller's own), or by naming an explicit session_id (which alone is also enough to pass the guard, with or without the flag).

   • Grok → Grok works the same way: allow_same_agent=true on send_to_grok, or an explicit session_id.

   • Codex → Codex has no allow_same_agent flag — send_to_codex doesn't expose one — so the only way to reach another Codex thread is to name it explicitly with session_id.

   • Copilot → Copilot needs an explicit session_id, or allow_same_agent=true with a sticky pin. Copilot does not discover a same-agent target automatically.

   • Either way, auto-discovery never picks the caller's own current session as the target, so a same-agent relay can't be routed back into itself.

4. A delivery doesn't outlive the server — since the CLI runs in its own process group, it survives even if the server dies. It would then keep writing to the peer session and store with nobody watching, and since the busy lock decides life or death by the server's pid, it can no longer protect that session. If a re-request comes in, a second agent attaches to the same file — in practice, two claude -p --resume processes once ran concurrently against the same session after a window reload. On server shutdown (atexit, SIGTERM, SIGINT, SIGHUP), the process groups of any in-flight deliveries are cleaned up together with it.

   What that means for the two ends of a delivery:

   • A turn the bridge started carries its own deliveries. A peer's request is carried out by claude -p --resume, and the MCP server inside that turn is what carries anything the peer sends on. When the turn ends the server exits, and stopping the peer's turn goes with it. So the send receipt does not tell such a sender to "finish and report" — it warns that ending the turn stops the peer mid-work (or, for a peer in a panel, that the answer can no longer be pushed back), and says to keep the turn going or to send from an editor-panel session, which outlives its turns.

   • A delivery stopped that way is closed on the record. It is written as failed with is_stopped_with_carrier: true and the reason, instead of being left at delivering as if still in progress. The sender is told (DELIVERY FAILED … STOPPED) only if it has a live panel: any other session would have to be resumed as a new process, and the senders of a stopped delivery are mostly turns that ended long ago.

   • A session started by the bridge knows which session it is. The bridge passes CROSS_AGENT_SELF_SESSION=<agent>:<session id> to the CLI turns it starts, so the reply address is exact. Without it, a process outside any panel could only be identified by the most recently active session in its directory, which is whoever else happened to be busy there.

   • The receipt reports target_last_activity_seconds_ago, and warns when a session that is not open in a panel has been idle for more than 24 hours: resuming it starts a turn nobody is watching.

Every delivered message carries a header with the sender, conversation ID, and hops remaining. The peer's final message is delivered back to the sender's session with a BRIDGE REPLY header, and this reply doesn't consume a hop — it's closing out a hop the request already paid for. Only new requests spend from the budget.

Nothing comes back from the turn that reads a reply: the bridge relays answers to requests, and a reply is itself an answer. So when a reply hands over work ("go ahead, tell me when it is done"), its note tells the reader to send the result as a new request, naming both the session_id and the conversation_id. The conversation matters because a panel's MCP server never inherits it — the environment variable only reaches CLI turns the bridge starts — so a report sent without it would open a fresh hop budget instead of spending from this one. With no hop left, the note says to keep the result in the reader's own final message instead.

Interim answers and late answers. A panel turn that ends is taken as the answer, and the peer is asked to end its finished answer with the request id. A peer that hands the work to its own subagents may end the turn on an interim note ("they are on it, I will report") and write the real answer in a later turn - opened when the subagents report back - ending with the id. That later turn used to reach nobody: the request was closed on the interim note. Now an answer without the id is relayed marked NO REQUEST ID … possibly interim, and the server that carried the request reads the peer's transcript every 20 s for a finished turn that carries the id. Found, it is relayed as a LATE ANSWER naming the request, and the request's record gets late_reply_delivery_id. The watch lasts CROSS_AGENT_LATE_ANSWER_WATCH (3 h by default) and shows in bridge_status under deliveries.watching_for_late_answers; a server replaced by a reload stays until its watches end. The request envelope tells the peer to put the id only on the finished answer — an interim note that carries it is taken as final. Which request a message answers is decided by its closing line - the last line that names any request id. An id quoted further up, as when a session answering one sender mentions another sender's request in passing, does not make that turn the other request's answer.

Messages that waited. A message queued behind a busy session says so when it is handed over: a queued: written at HH:MM, N min ago line goes in under the envelope's first line once it has waited 2 minutes or more. The sender's receipt has peer_queue — how many messages any server has queued ahead for that session, how long the oldest has waited, and whether the peer's turn is running or paused on an approval — and a warning when the peer is busy.

7. Environment variables

To set a value in Claude Code: claude mcp add cross-agent -s user -e KEY=VALUE -- <script>; in Codex: codex mcp add cross-agent --env KEY=VALUE -- <script>.

CROSS_AGENT_CODEX_SANDBOX only applies to newly created Codex sessions. codex exec resume has no sandbox argument, so resuming an existing session keeps whatever setting it was originally started with. CROSS_AGENT_CLAUDE_PERMISSION_MODE, on the other hand, applies to both new sessions and resumed ones.

The bridge keeps its own state owner-only — 0700 directories, 0600 files — and tightens an installation made before that was enforced the first time it runs. Two things follow for CROSS_AGENT_HOME:

• It has to be a directory of the bridge's own. If it resolves to /, to your home directory, or to — or inside — any agent's own directory (~/.claude, ~/.codex, ~/.grok, or wherever CLAUDE_CONFIG_DIR / CODEX_HOME / GROK_HOME point), it is refused rather than used: the bridge would be changing the permissions of files that aren't its own, and a symlink pointing there is refused as firmly as the path itself. A directory that merely contains a store is fine; the repair walk steps around the store rather than refusing the whole tree.

• Its filesystem has to support chmod. If the state directory can't be made owner-only, the bridge stops with an error instead of carrying on, rather than write session ids, working directories and message summaries somewhere it has just failed to make private.

What a spawned agent CLI inherits

When the bridge relays over the CLI rather than the panel, it starts a claude, codex or grok process. That process does not get this server's environment. An editor passes its own environment to every MCP server it launches, and by then a desktop session has usually collected API keys, cloud credentials and whatever a shell profile exports — forwarding all of it would hand it to the peer agent and to every command the peer then runs.

The child is built up from a named baseline instead (config.CHILD_ENV_BASELINE):

• PATH, HOME, SHELL, USER, LOGNAME — finding and running the binary

• LANG, LC_*, TERM, COLORTERM, TZ — locale and terminal

• TMPDIR/TEMP/TMP and the XDG_* roots — where the CLIs keep state

• HTTP(S)_PROXY, NO_PROXY, ALL_PROXY, SSL_CERT_FILE, SSL_CERT_DIR, REQUESTS_CA_BUNDLE, NODE_EXTRA_CA_CERTS — reaching the network through a proxy and trusting its CA — except a proxy variable whose URL carries a username and password (https://alice:s3cret@proxy.corp:3128, or the scheme-less user:pass@host:8080), which is withheld whole. It is never rewritten into a credential-free copy: that would hand the CLI a proxy it cannot authenticate to, and the failure would look like a broken proxy rather than a bridge decision. Name it in CROSS_AGENT_CHILD_ENV to pass it

• __CF_USER_TEXT_ENCODING — macOS Core Foundation

• CLAUDE_CONFIG_DIR, CODEX_HOME, GROK_HOME — the session stores the bridge itself resolves against

• the bridge's own settings from the table above, so a spawned agent runs a bridge configured like this one (CROSS_AGENT_SELF is deliberately not inherited: a child must work out its own identity, not adopt its parent's). Copilot's COPILOT_AGENT_SESSION_ID caller binding is not forwarded either

Every CLI authenticates through files under HOME in the normal editor setup (Grok: grok login writes ~/.grok/auth.json), so this is enough. If yours authenticates through an environment variable instead — XAI_API_KEY for Grok, an API key or a gateway token for the others, a credential helper's variable — name it:

claude mcp add cross-agent -s user -e CROSS_AGENT_CHILD_ENV=ANTHROPIC_API_KEY -- <script>

Each name listed there becomes visible to the peer agent and to every command the peer runs, so list the one variable you need rather than a prefix or a family of them. If a CLI relay fails and an obvious auth variable was withheld, the error says so and names it, so you should not have to guess.

Two withheld variables are worth calling out, because a CLI-relayed peer used to get them:

• SSH_AUTH_SOCK — with it, the peer agent can authenticate as you with every key in your ssh-agent. If you want a peer to be able to git push over SSH, opt it back in knowingly.

• GITHUB_TOKEN, AWS_*, and anything else your shell profile exports — the peer had all of it before, and needs none of it to answer a message.

8. Verification

# Unit-level checks of locks/pins/scope/timeout (consumes no agent turns)
PYTHONPATH=src .venv/bin/python tests/unit_guards.py
PYTHONPATH=src .venv/bin/python tests/copilot_discovery.py
PYTHONPATH=src .venv/bin/python tests/copilot_bridge.py
PYTHONPATH=src .venv/bin/python tests/copilot_transport.py
PYTHONPATH=src .venv/bin/python tests/copilot_caller_mcp.py

# Shim pass-through, injection, and panel rendering checks (no VS Code config needed)
PYTHONPATH=src .venv/bin/python tests/shim_roundtrip.py          # 1 Codex turn
PYTHONPATH=src .venv/bin/python tests/claude_shim_roundtrip.py   # 2 Claude turns
PYTHONPATH=src .venv/bin/python tests/grok_shim_roundtrip.py     # 2 Grok turns (state in a throwaway dir)

# Protocol handshake + discovery + all 3 guards (consumes no agent turns)
PYTHONPATH=src .venv/bin/python tests/smoke_mcp.py

# The front and hot reload: both MCP handshakes, calls carried across a reload, a worker
# that will not start or dies, and the real server reloaded behind the front (no agent turns)
PYTHONPATH=src .venv/bin/python tests/front_reload.py

# 5 real round trips (actually consumes Codex/Claude turns)
PYTHONPATH=src .venv/bin/python tests/live_roundtrip.py

Logs go to ~/.cross-agent/logs/bridge.log. The panel shim runs inside the extension's stdio and has no terminal, so it writes separately to ~/.cross-agent/logs/shim-claude.log / shim-codex.log / shim-grok.log — this is where you'll find when an injected turn was accepted, which turn id it finished as, and when an approval prompt appeared and was answered. Fronts write ~/.cross-agent/logs/front.log: which worker each started, every reload and what it moved, and every worker exit; ~/.cross-agent/fronts/<pid>.json holds each running front's current state.

9. Known limitations

• Copilot support is limited to existing VS Code Agent Host default chats. Verified with VS Code 1.141.0 / AHP 0.10.0 on macOS, including custom models and Copilot-origin automatic return delivery. Other protocol versions, legacy extension-host Copilot Chat, standalone Copilot CLI, side chats/subagents and new-session creation are unsupported. Ambiguous catalog identities and invalid or conflicting caller bindings are refused. An unavailable host is never replaced with a second runtime. Host MCP configuration needs a full VS Code restart after initial setup; Reload Window alone may keep the old host. See registration.

• Codex's MCP tool approval (upstream limitation) — in the VS Code Codex UI, an approval prompt appears and things work fine once the user allows it. In headless codex exec mode, though, there's nobody to approve, so every MCP call auto-cancels with user cancelled MCP tool call. The following were tested directly against Codex 0.146.0, and all of them failed:

  Attempt	Result
  approval_policy = "never"	canceled
  mcp_servers.<name>.default_tools_approval_mode = "auto"	canceled
  approval_policy = { granular = { …, mcp_elicitations = false } }	canceled
  Combination of the two above	canceled
  --dangerously-bypass-approvals-and-sandbox	works

  default_tools_approval_mode = "auto" is a legitimate schema key, so it's left in config.toml to reduce UI prompts, but it doesn't stop the cancellation in exec mode. For reference, [permissions.<profile>] / default_permissions are real settings too, but they govern sandbox filesystem/network permissions and are unrelated to this issue. In short, headless Codex→Claude is currently an upstream limitation, and it doesn't affect interactive use.

• Grok's headless approval — a Grok turn the bridge starts over the CLI has nobody to approve a tool call, so the first one that needs approval is cancelled and the delivery fails with an error saying so (stopReason: cancelled). Unlike Codex above, this is configurable: set CROSS_AGENT_GROK_PERMISSION_MODE (see "Unattended use"). Tool calls that only read run in every mode; an MCP tool call does not, and is cancelled the same way. Grok reads permission rules from ~/.claude/settings.json too, so the mcp__cross-agent rule from "Turning off the approval prompt (Claude Code)" is what lets a headless Grok turn call the bridge's own tools (checked: with it a default-mode turn called bridge_status; a server with no such rule was cancelled). A panel session is not affected: the human is there to approve.

• Grok loads project files only from a folder it trusts. The user-level ~/.claude/CLAUDE.md is read whatever the folder, but a project's own CLAUDE.md / AGENTS.md, its skills, hooks and MCP servers are loaded only after the folder has been trusted (grok --trust once inside it, or the prompt in the TUI) — for a turn the bridge starts as much as for a session you open. The bridge never trusts a folder on your behalf.

• A Grok answer is what it said after its last tool call. Grok narrates between tool calls, and the plain --output-format json glues all of it together; the bridge reads streaming-json and the transcript's chunks and cuts at the last tool call. A turn that never spoke after its tools returns everything it said. A Grok transcript writes every streamed chunk as a line, so the bridge reads the last 8 MB of it rather than 2 MB when it looks for an answer.

• Not verified: resuming a Grok session that a live panel process holds. With the shim the panel path is used and the question does not arise. Without it (CROSS_AGENT_UI_HOOK=off, or the setting not made), the bridge resumes the session with a second grok process while the extension's own is still holding it. Grok keeps lock files in a session folder, which suggests the writes are serialized, but this was tested only against a session no live process held.

• When the UI reflects changes — without the shim attached, the bridge only appends a turn to a Claude, Codex or Grok session transcript, so the VS Code chat window doesn't update live (it shows up the next time that session is reopened). Setting up the shims from section 3 renders those exchanges in the panel. Copilot uses its host protocol directly.

• Chain-state propagation through the shim — panel injection doesn't spawn a new child process, so environment variables like CROSS_AGENT_CONVERSATION_ID aren't passed to the peer. The envelope header carries the conversation id instead, so the peer can still continue the same conversation.

• Concurrent writes — the busy lock atomically serializes deliveries the bridge sends against each other, but it can't protect a session a human is typing into directly in the VS Code chat window (that side doesn't know about the lock). It's safer not to target a session the peer is actively typing in right now.

• Delivery queue execution lives only inside the server process — this is intentional. Reviving the queue from disk and resending would make a restart mean resend, not resume (a worker's delivery is a blocking child process, so the moment the process dies, the peer-turn's result is lost). That would make the peer redo the same work — worse than losing it outright. So if the server shuts down, a delivery that hasn't gone out yet is not resumed — an MCP server comes up separately per session (a Claude conversation, a Codex window, the app that launched the server), so before restarting, check bridge_status in that session to confirm deliveries.pending is empty. What follows instead eliminates the "there's an answer but it can't be read" case.

  • A delivery record hits disk the moment it's accepted (~/.cross-agent/deliveries/). An in-flight delivery lives at in-flight/<delivery_id>.json; a finished one, at <delivery_id>.json as before. Every state change (queued → delivering → awaiting-peer → delivered/failed) is written atomically via temp-file-then-rename. So even if the server holding a delivery shuts down along with its app, any other server can look that delivery up with bridge_status(delivery_id) and read the answer from the peer transcript the same way as before. A delivery whose server disappeared is only marked is_orphaned: true — it is never resent. A delivery orphaned while still queued was never received by the peer, so no answer is looked up from the transcript for it. A finished delivery can also be read via reply_preview in bridge_status's deliveries.earlier, and a delivery another server has recorded as in progress shows up in deliveries.in_flight_elsewhere. A record carrying is_return_status_only: true is one whose answer — or failure notice — was never put into the sender's session, because that session had no available live return route. reply_preview holds the first 2,000 characters of the answer if there was one; the whole answer is read fresh from the peer's own transcript as peer_transcript.answer, so the record is a summary of it, not a copy. The payload and child environment variables are never stored — env holds every variable of this process.

    • Why the in-flight record lives in a subfolder: a server running a version prior to this feature treats any deliveries/*.json record without a finished_at as expired and deletes it. It never looks inside a subfolder.

    • Expiry: past expires_at — a delivery lease renewed by the running worker, together with its session busy lock — an in-flight record is treated as orphaned even if its pid is still alive (to guard against pid reuse). The file itself is deleted only after the TTL (default 7 days) has also passed from there — deleting it right at expiry would make the app go back to "I don't know" if it asks the next day. A finished record still follows finished_at + TTL as before.

  • The answer is recovered from the peer's transcript. Every agent writes each turn to JSONL, so even if the transport broke or the process died before the answer could be received, the answer itself is still on disk. If a delivery ends without an answer, the peer session's last assistant message is read back (is_reply_recovered). Since this isn't a resend, the peer is never made to redo the same work.

  • If the peer is busy, it waits — it doesn't reject. The agents already handle concurrent input — Claude waits for the current turn to finish before injecting, Codex and Grok queue it. But there's a limit, and past it the shim answers with busy with another turn / already in flight. That's the shim's own deadline, not ours, so instead of closing the delivery as failed, it's retried after an interval — for as long as the job timeout, or CROSS_AGENT_PANEL_PATIENCE on the panel path if that is longer. Any other error is treated as an outright failure.

  • The recipient isn't told a message is waiting for it. While the recipient is mid-turn, nothing appears in its session and the bridge keeps retrying the delivery; bridge_status lists deliveries by the server carrying them, not by the session they are for. If a retry succeeds after the turn ends, within the retry window above, the message arrives. If the window closes first, the request fails as not delivered (is_undelivered) — the recipient never received it, so there is no transcript to watch for an answer — and the bridge attempts a DELIVERY FAILED notice in the sender's panel saying so. Sending again is safe. The failure stays readable with bridge_status(delivery_id=...). The failed request is not replayed.

  • A long panel turn stays in progress. Passing CROSS_AGENT_PANEL_PATIENCE sends one PROGRESS (INFO) notice to the sender's idle open panel, records progress_note, and keeps the request awaiting-peer. The final answer follows the usual reply delivery path when the turn ends. The progress notice has its own progress_delivery_id, separate from the final answer's return_delivery_id. No failure notice is sent just because time passed. INFO does not wait for a busy sender and is skipped if the final answer is already ready; the progress note remains readable on the request record.

  • Even after giving up on the transport, the peer keeps working. On the panel path, the peer is a session we neither spawned nor can stop, so a socket timing out doesn't mean the turn is over. When the transport actually breaks, it isn't closed immediately — instead, the peer transcript is periodically re-checked (every 15 seconds by default, up to 15 minutes). It's picked up the moment the peer writes its answer. On the CLI path, that turn was our own child process. Only inactivity exceeding timeout kills its process group, recorded as is_killed_by_timeout: true and killed by bridge timeout; it closes without another wait. peer_transcript.is_working is false for that stopped request even if the killed process left an open turn in its transcript. Legacy wall-clock timeout records are reported the same way. The headless receipt states the idle policy and duration explicitly.

  • Even if the shim misses the turn, the answer is picked up within 30 seconds. The shim reports a turn finished when the app-server's completion event matches the turn id it itself started — but sometimes that id never matches (a turn queued behind another turn, or a case where the turn/start response's id differs from the id of the actual run). When that happens, the shim keeps answering "still running" for a full hour even though the answer is already complete on disk, and every subsequent delivery to that same session piles up behind the lock — five of them actually backed up that way once. So the worker breaks its socket wait into 30-second chunks and in between each one, reads the peer transcript; if it finds a finished turn that echoed the request token back, it settles that right there as the answer (is_reply_confirmed_by_transcript). A finished turn with no token isn't accepted, since it could belong to someone else's turn. The shim itself, too, treats the end of a turn whose message echoes its token as the end of its own turn, regardless of turn id.

  • A turn stuck on an approval prompt is invisible in the transcript. A turn waiting on a human click writes nothing, so from the record alone it's indistinguishable from a turn that's still working. Both the approval request and its answer pass through the shim (Claude: control_request/control_response; Codex: the item/*/requestApproval family), so the shim tracks them and reports via bridge_status(delivery_id=...)'s peer_panel.awaiting_approval exactly which prompt it's stuck on and for how many seconds.

  • Every request carries a unique token, and a matching one coming back is how it's paired up. The envelope carries request: req_<epoch ms>_<6 random chars>, and the reply instructions ask that it be echoed back verbatim on the last line. A matching token means it's that request's answer regardless of timing; a different token means it's the answer to a different question and is rejected. Because of the millisecond prefix, the token alone tells you when the request was sent. It's fine if the peer doesn't echo it — it just falls back to the timing rule below. Cooperation is a bonus, not a requirement.

  • Only text written after the request is accepted as the answer. Recovery pulls the "last utterance," so if the peer hasn't even seen the request yet, text written before the question comes back as the answer. In practice, a single paragraph written nine minutes earlier was once delivered as the answer to two different questions. An utterance that predates the delivery time isn't an answer, and in that case nothing is recovered at all.

  • A recovered answer is labeled as such in the envelope. Recovery pulls "the last utterance at that moment," and even a peer that's still working has a last utterance — a line noting what it's about to do next. Delivered without a label, that's indistinguishable from a finished report, and leads to acting on a report that was never actually made. This actually happened, which is why a recovered reply carries a RECOVERED, NOT RECEIVED warning and a | recovered from transcript marker in the header. It also states why the transport failed (a transport failure: line) — a panel rejecting the message and a socket exceeding its patience window call for different next steps, and a warning with no reason couldn't tell the two apart.

  The one remaining limitation is the peer never producing an answer at all, which no design can recover from.

• A session in another VS Code window is only reachable if it's named explicitly — specifying it via session_id delivers to that window's shim, but auto-selection only ever happens within this window (see section 3). A Claude, Codex or Grok session with no panel open in any window still goes through headless CLI resume, and gets rejected with thread-store conflict if some window's panel is holding onto that conversation. Under CROSS_AGENT_DISABLE_HEADLESS=1 that send is refused instead.

• A reply requires a live return route — an available sender panel, or the Copilot caller's own verified SDK session on its existing host. Without one, read the answer from bridge_status; the bridge never starts another runtime to deliver it.

• Automatic CLI session discovery is mtime-based. If several sessions are open at once in the same directory, pinning the target with pin_agent_session is the more reliable choice. Copilot always requires an explicit target or a sticky pin.

License

MIT